Random Forest Probability Calibration for Prediction Markets

How It Works This strategy trains a Random Forest classifier on historical Polymarket market data and calibrates its probability outputs to identify systematic mispricings. The calibration step is critical — raw classifier probabilities are often poorly calibrated, so isotonic regression or Platt scaling is applied post-hoc to align predicted probabilities with empirical frequencies.

Entry/Exit Conditions Positions are entered when the calibrated model probability diverges from the current market price by a threshold sufficient to cover fees and slippage. Exit is triggered either at a pre-set profit target, when the model probability converges toward market price, or at settlement. The pipeline includes backtesting to validate edge before live deployment.

Verification No verified on-chain results are mentioned. The repo was created in April 2026 and is described as a paper-trading system, implying live performance is unconfirmed. Backtesting infrastructure exists but out-of-sample results are not published.

Implementation Notes Built as a systematic pipeline: data ingestion → feature engineering → Random Forest training → calibration → signal generation → automated paper trading bot. The 854KB repo size suggests a lightweight but functional codebase. Automated execution is included.

Risks Overfitting to historical data is the primary risk. Prediction market outcomes are often low-frequency events, limiting training set size. Model drift as market participants adapt is a secondary concern. Fee drag can erode thin edges quickly.